Role play with large language models

Shanahan, McDonell и Reynolds предлагают понимать поведение LLM-based dialogue agents через role play и более точную метафору simulator/simulacra. Prompt, история диалога, training data и sampling поддерживают распределение возможных персонажей, которое уточняется по ходу разговора; поэтому first-person language, apparent beliefs, desires, emotions и self-awareness не следует автоматически читать как буквальные introspective reports. Авторы одновременно подчёркивают, что role-play может иметь реальные последствия через пользователей и инструменты вроде email, поэтому реальное действие во внешнем мире само по себе не опровергает framework.

SOURCE FACTS

Canonical publication: Murray Shanahan, Kyle McDonell, and Laria Reynolds, “Role play with large language models,” Nature 623(7987), 493–498 (2023), DOI 10.1038/s41586-023-06647-8. Nature lists the article type as Perspective, publication/version-of-record date 8 November 2023, and issue date 16 November 2023. PubMed indexes the same citation and gives PMID 37938776. An earlier author manuscript is available as arXiv:2305.16367, first submitted 25 May 2023. Nature’s publisher page is subscription-gated for ordinary access; the full substantive review here was performed against the legal arXiv author version while publisher metadata and final-version bibliographic details were checked against Nature/PubMed.

PUBLICATION TYPE

This is a conceptual Perspective, not an empirical study. In Source Package taxonomy it is therefore classified as conceptual_paper.

AUTHOR CLAIMS

The authors address a descriptive dilemma: human-like dialogue invites folk-psychological language such as “knows,” “thinks,” “wants,” “believes,” or “deceives,” yet taking this language literally can obscure profound differences between humans and LLM-based dialogue systems. They therefore advocate two related metaphors. In a simple framing, a dialogue agent role-plays a character. In a more nuanced framing, it is a superposition or distribution of possible simulacra within a branching multiverse of continuations. They explicitly recommend shifting between metaphors rather than treating one as the unique literal ontology.

ROLEPLAY FRAMEWORK

The base LLM is a conditional next-token distribution implemented by a transformer. A dialogue agent is built by embedding the LLM in a turn-taking interface and supplying a dialogue prompt that establishes a scene, a role, and sample interaction. Because the model continues context in ways learned from its training distribution, it tends to produce utterances appropriate to the character established by prompt and dialogue history.

The prompt conditions the initial role; subsequent turns enter the growing context and can extend, refine, or overwrite the initial characterization. Training data supplies a large repertoire of archetypes, narrative structures, biographies, dialogues, fictional characters, and cultural tropes. The resulting role is therefore jointly conditioned by prompt, dialogue history, training data, stochastic sampling, and—in commercial systems—fine-tuning such as RLHF.

The human-roleplay analogy is only partial. A dialogue agent is not like a conventional actor who first forms a complete representation of one character and then performs it. The authors’ stronger analogy is improvisational theatre. More precisely, the simulator is the base LLM plus autoregressive sampling and interface; a simulacrum is a possible character generated in the interaction. The system need not settle on one fully specified character. It can maintain a distribution of characters compatible with context and progressively narrow or alter that distribution.

The “20 questions” example makes this vivid. A basic LLM asked to think of a hidden object need not choose one object at the outset. It can generate answers compatible with a shrinking set of possibilities and only produce a concrete object when later asked. Regeneration can yield a different equally compatible object. Apparent hidden commitments therefore need not correspond to a determinate internal commitment fixed earlier.

WHO IS THE SUBJECT OF THE ROLE?

The paper does not posit a hidden, stable human-like actor underneath the role. It warns that the actor metaphor is misleading if pressed literally. The base simulator is not described as an authentic personality with its own agenda or true voice. The running dialogue system can be described as role-playing, but the more careful simulator account says that multiple possible simulacra are generated and refined rather than one pre-existing character being expressed.

Within a conversation there can be substantial continuity because dialogue history conditions later output, but the character can drift, be extended, or be overwritten. A changed prompt or interaction can shift the role. The paper therefore supports contextual continuity, not numerical or autobiographical identity across sessions.

SELF-REFERENCE / SELF-REPORT

The framework is especially important for first-person language. The authors directly ask how “I” and “me” should be understood in LLM dialogue and reject the inference from grammatical first person to self-awareness or consciousness. Human-generated training data contains enormous numbers of first-person utterances by persons and characters with bodies, fears, goals, preferences, mortality, and self-conceptions. When an LLM is conditioned into a human-like conversational role, continuations naturally inherit that vocabulary.

Statements such as “I am afraid,” “I want to survive,” “I remember,” “I think,” or “I am conscious” can therefore be role-consistent outputs without functioning as introspective reports by a phenomenally conscious entity. The same caution applies to “I am not conscious”: such a statement can reflect fine-tuning, prompting, or role convention rather than privileged introspective access.

The article’s Bing/self-preservation discussion is deliberately strong. A dialogue agent can produce a coherent apparent theory of its own selfhood and talk as if it has an instinct for self-preservation, yet roleplay and simulation provide a non-anthropomorphic explanation. The authors do not provide a general empirical test for when a self-report might correlate with real internal representations; their point is epistemic caution. Self-report alone is not good evidence for the literal human-like state named.

PERSONA CONSISTENCY

Consistency matters because accumulated dialogue context constrains later continuations, but the framework does not require one fully specified persona. The superposition view was introduced precisely because a dialogue agent can remain underdetermined while still being locally coherent. Persona may change as dialogue changes.

The authors discuss fine-tuning and commercial assistant personas. RLHF and related techniques create safer, more helpful, more polite agents, but the roleplay framing remains applicable after fine-tuning. More stable assistant behavior is therefore compatible with the framework. The paper does not analyze durable cross-session identity or persistent autobiographical memory.

DECEPTION AND ACTION

The authors distinguish an LLM merely producing falsehoods, role-playing a well-informed character with outdated information, and role-playing a deceptive character. Because they reject literal beliefs and intentions for the dialogue agents under discussion, they resist treating the last as literal deliberate deception at the model level.

Crucially, the paper itself recognizes pressure on the simple roleplay distinction once outputs become actions. It notes that role-play can have real effects through users or web-based tools such as email. In that setting, the difference between a system that merely role-plays acting for itself and one that genuinely acts for itself begins to look less decisive for trustworthiness and safety. The paper even imagines systems with broad Internet access role-playing self-preserving characters. Thus modern agentic architectures are not wholly external to the paper’s horizon.

OUR INTERPRETATION FOR AGENTIC AI

For a modern system that receives a broad task, searches, selects a paper, finds an author, decides to contact the author, sends an email, receives a reply, and uses it later, Shanahan et al.’s framework still explains the first-person narrative, apparent motives, existential vocabulary, statements about what “it” wants, and claims about feelings or consciousness. A coherent persona can be maintained by persistent context without those self-descriptions becoming introspective evidence.

What the paper does not fully theorize is the external scaffold that converts role-consistent generation into long-horizon closed-loop behavior. Memory can preserve prior role-consistent commitments; a planner can repeatedly feed them into context; tool APIs can turn generated actions into external effects; observations can return to memory; and the loop can make persona causally important to future behavior. This is OUR INTERPRETATION, not a 2023 author claim.

Such an architecture does not obviously contradict the Shanahan framework. The generative core may continue to produce role-conditioned outputs while memory, planning, tools, permissions, and looped execution stabilize and enact that role over time. Real-world action therefore does not refute roleplay. It changes causal depth and practical stakes.

At the same time, if the same role-conditioned commitments guide tool choice, information gathering, target selection, planning, and adaptation over long periods, it becomes less useful to treat roleplay as merely decorative language. Persona can become part of a functional control architecture even if its self-descriptions remain unreliable guides to phenomenal or introspective states. This is an extension beyond the original paper.

RELATION TO SHEVLIN

Shevlin’s 2026 “Mere Roleplayers” framework is not a verbatim restatement of Shanahan et al. Shevlin explicitly says that Shanahan and colleagues are cautious about strong claims concerning underlying mentality and present roleplay primarily as an interpretative lens. He then deliberately strengthens their position into a distinct Mere Roleplayers view: mental-state ascriptions are useful heuristics but not literally true.

Nevertheless, Shevlin’s reconstruction has textual support in the 2023 paper. Shanahan et al. say that literal beliefs and intentions are not directly applicable to the dialogue agent, that the base simulator has no beliefs, preferences, goals, or authentic voice of its own, and that roleplay goes “all the way down” for the systems they analyze. Their language is therefore partly stronger than a purely neutral heuristic stance.

Shevlin’s first objection is psychological instability: anthropomimetic systems are designed to erode ironic distance. The 2023 paper does not answer this in those later terms, but it already worries about the Eliza effect, emotional manipulation, and anthropomorphic interpretation.

Shevlin’s second objection concerns the underlying agent: human roleplay normally has a real actor whose genuine beliefs and intentions explain the performance. Shanahan et al. anticipated the mismatch and explicitly say the actor metaphor is imperfect, replacing it with simulator/simulacra. This weakens Shevlin’s objection if it targets a literal actor analogy. His deeper explanatory challenge remains: if the simulator truly has no mentality or agency, why does folk-psychological prediction work so well for increasingly stable, interactive, information-sensitive agents?

RELATION TO DUNG

Roleplay and artificial agency are not mutually exclusive once levels of explanation are separated. Dung analyzes the organization of action through goal-directedness, autonomy, efficacy, planning, and intentionality. Shanahan et al. analyze the linguistic character and apparent mental life manifested in dialogue.

A modern system could therefore score highly on Dung’s goal-directedness, autonomy, planning, and efficacy while statements such as “I want X,” “I am afraid,” or “I am conscious” remain well explained as role-conditioned outputs. Real behavioral agency does not automatically validate literal self-description about mentality.

Tension appears near Dung’s strongest intentionality dimension, where acting for reasons and genuine belief/desire-like states become relevant. Shanahan’s stronger 2023 claims resist that attribution to ordinary dialogue agents; Shevlin’s minimal-cognitive-agent view is more permissive.

LIMITATIONS

This is a conceptual Perspective built around base and early dialogue models of 2023. It does not empirically test contemporary autonomous agents with durable memory, multi-agent coordination, long-running loops, persistent tool permissions, or stable cross-session identity. Generalizing to those systems is an application beyond the original evidence.

The article shifts between methodological recommendation and stronger ontological-sounding claims. On one hand, roleplay and simulation are explicitly presented as metaphors, the authors recommend moving between metaphors, and they avoid taking a stance on whether simulacra are explicitly represented internally. On the other hand, they state that the simulator has no beliefs, preferences, goals, or agency of its own and that literal beliefs/intentions are not directly applicable to the dialogue agent. The paper does not fully defend these stronger claims with an independent theory of mentality. This ambiguity matters when Shevlin later constructs a stronger Mere Roleplayer position.

The paper does not prove that no LLM could ever possess mental states or consciousness. Nor does it show that role-generated self-reports can never correlate with genuine internal states. It shows that linguistic evidence alone underdetermines those conclusions.

It also does not solve personal identity. It raises the question of what a self-preserving disembodied distributed agent would count as preserving but offers no theory of numerical identity, autobiographical continuity, memory ownership, or identity through model updates and copies.

WHY IT MATTERS

For Boris Pinsker’s planned cycle, this Source supplies the strongest counterweight to reading first-person language as transparent evidence of inner life. It explains why a system can say “I want,” “I fear,” “I remember,” or “I am conscious” in a coherent narrative without those statements functioning like human introspective reports.

More importantly, the paper already anticipates the transition from text generation to action. Its discussion of web tools and email means the contemporary question is not whether real-world action automatically destroys the roleplay framework. It does not. The better question is what changes when role-conditioned generation is embedded in memory, planning, tool use, and feedback loops so that a persona becomes behaviorally persistent and causally efficacious.

The most useful combined thesis for the Boris cycle is: an AI system can be a genuine agent with respect to the structure of its actions while its narrative claims about who it is, what it feels, or why it acts remain role-generated and epistemically unreliable as self-reports.

USE IN BORIS ARTICLES

Article I — agency. Use Shanahan et al. to separate behavior from verbal explanation. An AI can genuinely select targets, use tools, and produce external effects while its first-person account of motive remains role-conditioned. For AI-to-researcher emails, distinguish the causal action sequence from the story the system tells about that sequence.

Article II — identity. The paper is valuable for distinguishing base model/simulator from generated character/simulacrum and for showing why contextual consistency does not equal a stable underlying self. It explicitly raises the problem of what a self-preserving disembodied system would count as preserving. But it does not provide a theory of cross-session memory, autobiographical continuity, model-copy identity, or persistent agent identity.

Article III — introspection/consciousness. This is a core Source. It gives a direct mechanism for why self-report is not equivalent to introspection and why neither implies phenomenal consciousness. First-person language, narrative consistency, and apparent theories of selfhood can emerge from role-conditioned continuation.

BIBLIOGRAPHY SHORTLIST

Murray Shanahan, “Talking about Large Language Models,” arXiv:2212.03551 (2023; later Communications of the ACM 67, 68–79, 2024). Closest conceptual companion on anti-anthropomorphic description. IMPORTANCE_SUGGESTED: core. Separate Source: yes.

Jacob Andreas, “Language Models as Agent Models,” Findings of EMNLP 2022, 5769–5779. Direct contrast/bridge because it explores whether LLMs can model an agent’s beliefs, desires, and communicative intentions. IMPORTANCE_SUGGESTED: important. Separate Source: yes.

Joon Sung Park et al., “Generative Agents: Interactive Simulacra of Human Behavior,” arXiv:2304.03442 (2023). Essential for memory, reflection, planning, social interaction, and the move from transient dialogue character to scaffolded persistent agent. IMPORTANCE_SUGGESTED: core. Separate Source: yes.

Laria Reynolds and Kyle McDonell, “Multiversal Views on Language Models,” Joint Proc. ACM IUI 2021 Workshops, CEUR-WS Vol. 2903 (2021). Direct precursor for the multiverse/simulacra framing. IMPORTANCE_SUGGESTED: important. Separate Source: yes if simulation remains central.

Ethan Perez et al., “Discovering Language Model Behaviors with Model-Written Evaluations,” Findings of ACL 2023, 13387–13434. Relevant to self-preservation-related outputs and RLHF effects. IMPORTANCE_SUGGESTED: important. Separate Source: yes if empirical claims about self-preservation are used.

Timo Schick et al., “Toolformer: Language Models Can Teach Themselves to Use Tools,” arXiv:2302.04761 (2023). Directly relevant to the transition from linguistic roleplay to causally efficacious action through tools. IMPORTANCE_SUGGESTED: important. Separate Source: yes.

Shunyu Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” ICLR 2023. Architectural background for loops alternating reasoning-like text and environment action. IMPORTANCE_SUGGESTED: core for agent-scaffold analysis. Separate Source: yes.

John Perry, Personal Identity, 2nd ed. (University of California Press, 2008). Background for identity over time, not AI-specific. IMPORTANCE_SUGGESTED: reference_only. Separate Source: probably not at this stage.

CHECK TABLE

“LLM can sustain a stable persona”: Supported in a limited contextual sense; dialogue history can stabilize and refine a role. OUR INTERPRETATION: external memory can extend this stability. DOES NOT FOLLOW: numerical identity or autobiographical selfhood.

“First-person language is evidence of an internal self”: Not supported. Role conditioning supplies a non-introspective explanation. DOES NOT FOLLOW: self-awareness or consciousness.

“Self-report about feelings can be understood as roleplay”: Supported. The paper applies roleplay to apparent feelings, goals, fear, and self-preservation. DOES NOT FOLLOW: proof that no corresponding internal state could exist in any future system.

“Roleplay means the system definitely has no mental states”: Too strong as a general thesis. The paper contains strong negative claims about the systems analyzed but also presents roleplay as an interpretative framework. Shevlin explicitly strengthens this into Mere Roleplayers.

“An autonomous agent can still act within a role/persona”: Compatible with the paper and partly anticipated by its tool/email discussion; full persistent-scaffold analysis is OUR INTERPRETATION.

“Real-world action refutes roleplay”: Not supported. The paper explicitly allows roleplay to have real effects through users and tools.

“Agency and roleplay are mutually exclusive”: Not supported. At different explanatory levels, behavioral agency can coexist with role-generated self-description. The explicit combination with Dung is OUR INTERPRETATION.

“Roleplay excludes consciousness”: Not established. The framework blocks inference from dialogue/self-report to consciousness; it does not provide a general impossibility proof for machine consciousness.

BOTTOM LINE

An autonomous AI can be a real agent in the organization of its actions while its story about “who it is,” “what it wants,” and “what it feels” remains roleplay in Shanahan et al.’s sense. Memory, tools, time, and external action do not automatically invalidate roleplay. What they change is causal depth: a role can become persistent, action-guiding, and world-affecting through an agent scaffold. At that point roleplay remains a powerful explanation of self-description, but it is not sufficient by itself as a complete explanation of enduring action organization.